Goodness of fit - structural equation modeling (GSEM)

[Gwinyai Masukume]

Greetings,

Using Stata 13:

When using the GSEM (generalized structural equation model) it would appear the options to test for goodness of fit are grayed out. How do I test my model(s) goodness of fit?

Is having negative ‘B’s’ appropriate?
I’m also estimating using ‘clustered robust’, would matrices of residuals be best to assess goodness of fit?

I am new to structural equation modeling.

Thanks and yours sincerely,
Gwinyai

[Richard Williams]

-help sem estat gof- says "estat gof is for use after sem but not gsem." I don't see why negative coefficients would be inappropriate.

[Gwinyai Masukume]

Many thanks Richard. Does this mean when using GSEM one does not test goodness of fit? A perhaps elementary question - what's the major difference between SEM and GSEM?

[Richard Williams]

http://www.stata.com/stata13/generalized-sem/

[Richard Williams]

If sem can meet your needs, use it. http://www.stata.com/stata12/structu...tion-modeling/

[Gwinyai Masukume]

Thanks.

[ben earnhart]

Sometimes a GSEM will allow a negative variance for some variance parameters; compared to Lisrel and Mplus, it is much more forgiving of what would otherwise be "Heywood cases." Which is good, because the imposition of a positive variance is a bit arbitrary when the variance itself is not directly computed using some sort of sum-squared formula. But it plays havoc with goodness-of-fit indices, which are predicated on positive variances. Negative variances are a non-issue when the software won't allow you to have them. But GSEM lets you make up your mind whether such anomalies are acceptable, and thus will not automatically dish out GFI type numbers.

Check out SEMnet, devoted to SEM, and you may get a better and more complete answer, but I hope this helps a bit as to why GSEM won't do GFIs, indirect effects, and some other things we take for granted.

[ben earnhart]

ps: http://www2.gsu.edu/~mkteer/semnet.html#Joining for taking your question to SEMNET. Yes, the Stata folks cross-breed with them to a large degree, but taking advantage of a SEM-specific community could help get you a good answer, and possible alternatives using Lisrel or Mplus.

[Gwinyai Masukume]

Thanks Ben.

[SRL]

I have another question related to gof
Is it possible to assess model fit when using vce(cluster clustervar) in sem?
Found this thread http://www.stata.com/statalist/archi.../msg00800.html but considering it's couple years old, and sem in state has improved quite a bit since then, I was wondering whether this is still the case.

Thanks

[Radhouene DOGGUI]

You may have look to the following paper: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC8087284/